Finally, ELIZA: A Platform, Not a ChatBot!

Table of Links

Abstract and 1. Introduction

2. Why ELIZA?
3. The Intelligence Engineers
4. Newell, Shaw, and Simon’s IPL Logic Theorist: The First True AIs
5. From IPL to SLIP and Lisp
6. A Critical Tangent into Gomoku
7. Interpretation is the Core of Intelligence
8. The Threads Come Together: Interpretation, Language, Lists, Graphs, and Recursion
9. Finally ELIZA: A Platform, Not a Chat Bot!
10. A Perfect Irony: A Lisp ELIZA Escapes and is Misinterpreted by the AI Community
11. Another Wave: A BASIC ELIZA turns the PC Generation on to AI
12. Conclusion: A certain danger lurks there
13. Acknowledgements and References

9 Finally ELIZA: A Platform, Not a Chat Bot!

In short order, after the Datamation article, Weizenbaum publishes a series of papers in the Communications of the ACM (CACM), the premier publication in computing research and development at that time. His SLIP paper appears in 1963[49], and just three years later, in 1966, having moved to Stanford and then MIT, all of this comes together into the ELIZA paper.[47]

Again, let’s carefully read what should be obvious, but is commonly overlooked, the title of the paper: “ELIZA– A Computer Program For the Study of Natural Language Communication Between Man And Machine.” This paper begins with a somewhat less negative version of the same plaint as the gomoku paper:

“Introduction. It is said that to explain is to explain away. This maxim is nowhere so well fulfilled as in the area of computer programming, especially in what is called heuristic programming and artificial intelligence. For in those realms machines are made to behave in wondrous ways, often sufficient to dazzle even the most experienced observer. But once a particular program is unmasked, once its inner workings are explained in language sufficiently plain to induce understanding, its magic crumbles away; it stands revealed as a mere collection of procedures, each quite comprehensible. The observer says to himself “I could have written that”. With that thought he moves the program in question from the shelf marked “intelligent”, to that reserved for curios, fit to be discussed only with people less enlightened than he.”[47, p. 36]

Weizenbaum continues in the next short paragraph: “The object of this paper is to cause just such a re-evaluation of the program about to be “explained”. Few programs ever needed it more.”[47, p. 36] This last sentence reveals something hidden: “Few programs needed it more.” This echo of the gomoku paper tells us that it is exceedingly likely that Weizenbaum had already had the experiences that become so famous much later, of people talking intimately to ELIZA, and which became so antithetical to Weizenbaum, leading him to his later positions on AI.

He only briefly revisits this later in the paper: “A large part of whatever elegance may be credited to ELIZA lies in the fact that ELIZA maintains the illusion of understanding with so little machinery.” But he immediately returns to technical matters “But, there are bounds on the extendability of ELIZA’s “understanding”...” And once more, the very last paragraph of the paper reads:

“The intent of the above remarks [about translating processors] is to further rob ELIZA of the aura of magic to which its application to psychological subject matter has to some extent contributed. Seen in the coldest possible light, ELIZA is a translating processor [...] which has been especially constructed to work well with natural language text.”

Interestingly, this is almost identical to Searle’s “Chinese Room” argument about AI, which appeared significantly later, in 1980.[40] If any program was ever to work in exactly the version described by Searle, it would be ELIZA! Indeed, Searle specifically mentions ELIZA: “I will consider the work of Roger Schank and his colleagues at Yale [...]. But nothing that follows depends upon the details of Schank’s programs. The same arguments would apply to Winograd’s SHRDLU [...], Weizenbaum’s ELIZA [...], and indeed any Turing machine simulation of human mental phenomena.” [italics added]

Let us pause to take note of how this last phrase of Searle’s ties together (under his critique, in this case) another set of threads, this time from Hilbert through Turing and up to ELIZA and its descendants. Under the present hypothesis, Searle and other critics of ELIZA as an AI miss the point of ELIZA; ELIZA was not intended as an AI (orchatbot) at all, but as a platform for experiments in human interaction with AIs, and in particular the problem of interpretation; a problem that is now much more important, and equally understudied as it was 60 years ago, when Joseph Weizenbaum set out to (by this hypothesis) build a platform for research in this area, even going to far as the describe ELIZA as a vehicle for running a version of Turing’s imitation game:

“With ELIZA as the basic vehicle, experiments may be set up in which the subjects find it credible to believe that the responses which appear on his typewriter are generated by a human sitting at a similar instrument in another room. How must the script be written in order to maintain the credibility of this idea over a long period of time? How can the performance of ELIZA be systematically degraded in order to achieve controlled and predictable thresholds of credibility in the subject? What, in all this, is the role of the initial instruction to the subject? On the other hand, suppose the subject is told he is communicating with a machine. What is he led to believe about the machine as a result of his conversational experience with it? Some subjects-have been very hard to convince that ELIZA (with its present script) is not human. This is a striking form of Turing’s test. What experimental design would make it more nearly rigorous and airtight?”[47, p. 42]

Understanding ELIZA as a platform for research into human-machine communication is a very different framing of ELIZA than is commonly attributed to Weizenbaum. On the one hand, in his 1975 book[46] Weizenbaum frames AI as dangerous and/or immoral, and mentions ELIZA very little. His papers circa ELIZA, around the mid 1960s, as we saw above, describe a weaker, although related framing of ELIZA and other AIs of the time as intentional illusions.

But there is a third way of framing ELIZA that aligns both of these, and finally draws together all of the threads that we have discussed. This third framing might be rendered as: “AI could be dangerous, and is possibly immoral, if people interpret it the wrong way. Therefore, we need to study how people interpret their interaction with complex computer programs, especially ones that may appear to be intelligent. ELIZA is a platform in support of that project.” Under this theory ELIZA is not an AI at all, and is only barely a chatbot. Rather, ELIZA is a platform for research into how human interpretation works, and potentially how it may go awray or be abused. Indeed, Weizenbaum himself wrote a detailed outline of a paper discussing potential experiments along these lines that was recently uncovered in Weizenbaum’s archives at MIT.[2]. That paper, signed in Weizenbaum’s own hand, describes experiments he envisioned to be carried out with ELIZA, including detailed discussion of potential experiments in discourse and discourse correction. The very first page of that outline, after a brief introduction, dives right into the heart of the questions we have described in the present paper as the ones that Weizenbaum was most interested in, and which ELIZA was designed to explore:

“Understanding and misunderstanding. A. Our concern with partial understanding and thus understanding and misunderstanding, in diadic communication derives from both the clinical observation and the experimentally demonstratable facts that one or both parties in a two-party communication can be under the impression that they are understood by the other, while, in fact all that is understood is but a fragment of what is said or typed.”

That outline was apparently intended to become a paper published alongside the ELIZA paper, but, sadly, as far as we know it never made it beyond the draft stage. And, unfortunately, Weizenbaum never explored the potential potential for ELIZA as an experimental platform. Instead, at least at MIT, ELIZA was primarily explored for its educational potential (also mentioned briefly on the above-described outline), a project carried forward for a time by Paul Hayward, Edmund Taylor, and Walter Daniels.[31, 43]

Although Weizenbaum never himself used ELIZA as a platform for research into human-machine interpretation, some researchers did. In particular, a trio of researchers working right across the Charles river from MIT, at The Stanley Cobb Laboratory for Psychiatric Research, of the Massachusetts General Hospital and Harvard Medical School, explored ELIZA as “a new research tool in psychology.”[34] They observed that ELIZA “allows the stabilization of one party in dyadic communication, for ongoing analysis of certain types of communication and for systematic hypothesis testing about such communication”[34, p. 190] Although these researcher were not explicit studying the interpretive aspects of human-computer communication that interested Weizenbaum, it came up regardless: “The plausible conversation which can be generated by these machines often leads to disputes over the alleged intelligence of the computer. Such disagreement may distract attention from the important contributions such a machine may make to the understanding of communication. For each computer program is a simplified model of dyadic communication and may greatly assist in theory construction and testing.”[38, p. 165]. These researchers were using Weizenbaum’s MAD-SLIP version of ELIZA,although with a slightly different script (called “YapYap” [10]) and recently researchers studying the archives of Harold Garfinkel discovered original transcripts of their subjects’ conversations with this ELIZA, and the YapYap script, and demonstrated that this script exactly recreated the original conversations.[10]

Garfinkel and his coworkers interests seem to have been more aligned with what Weizenbaum had in mind: “Garfinkel was interested in how human–computer interaction was exploiting human social interactional requirements in ways that not only forced participants to do the work of making sense of a chatbot’s turns, but also gave them the feeling of an authentic conversation.”[25] Unfortunately, Garfinkel’s research in this area was never published, and research with ELIZA, and indeed all research into the important interpretive phenomena that interested, and later horrified Weizenbaum, appears to have ended at this point.[12, 13]

[12] With the exception of the Hayward et al. educational research, mentioned previously.[31] This work was carried out on a highly modified ELIZA, and with an explicit educational goal, again not directly addressing the interpretive problem in human-computer communication.

[13] Garfinkel and his coworkers did take this topic up quite directly, and worked for a time with the MGH team, although the Garfinkel-related research ended up using a different platform, called LYRIC, for their work. And, again, was never published.[25]